Epoch: GPT-6.2 does 12.3 serial reasoning steps per token, raising monitorability concerns
austinc3301 · x · 2026-11-18
Epoch AI measured that GPT-6.2 Astra performs 12.3 serial reasoning steps in a single output token without chain of thought, versus 10.5 for GPT-6 Astra and 8.6 for GPT-6.1 Sol; before GPT-6 Astra, the best was Claude Fable 5.1 at 4.2, with Claude Opus 3 at 2.3. Steps are normalized for complexity at 25% above-chance accuracy. DKokotajlo reads this as OpenAI going deeper into the neuralese/recurrence tech tree and argues it must degrade monitorability — he finds it hard to believe OpenAI insiders who say recurrence doesn't matter much directly and it's all about pretraining loss. Models capable of complex reasoning without visible CoT are harder to supervise.
More from AGI Musings
- Gary Marcus shares 'famous last words' table of Enron, Lehman, SVB, FTX CEOs to mock AI hype — GaryMarcus · 2026-11-18
- A 26-Patent Inventor Asks: Can Curiosity Be Taught in the AGI Era? — tctjr · 2026-11-18
- AI equity researcher Chinasa Okolo named a Gates Foundation Goalkeepers leader — ChinasaTOkolo · 2026-11-18
- Microsoft AI: Enterprises don't want to pour their IP into someone else's model — JFPuget · 2026-11-18
- Essay: AI isn't the future — it will trap us in our own past — whoamisri · 2026-11-18
- Musk on AI Race: 'Being a Spectator or a Participant' — He Chose to Race — austinc3301 · 2026-11-18